Wednesday, February 3, 2016

T-Ranketology

***Revamp in 2018!

As you'll see below, the original method of T-Ranketology was to take five ranks—T-Rank, Elo, WAB, RPI, and "Resume" rank—and given them various weights to come up with a score (with a couple other tweaks like a good record bonus and a bad record penalty).

Recently I wanted to come up with a way to use the T-Ranketology score to give an estimate of a team's chances for receiving an at-large. In looking into this, it occurred to me that I could use the five inputs to do a "logistic regression," and use the resulting model to provide the chance I was looking for, using the actual ranks themselves (rather than my own weighted score) as the input.

I did this, and the results were pleasing. So pleasing that I've now decided to abandon the old "weighted score" method and just use the model from the logistic regression all the way down. The results are pretty similar—after all, the inputs are the same—but I can do a bit more with the new system, and feel a bit more confident in the results. For example, on Teamcast I can now provide an estimate of how much a win or loss affects a team's chances, in percentage terms, of making the tournament.

So there you have it, that's the T-Ranketology update for this year.

***Another update:

See this post for some changes made in 2017.

***Update:

See this post for an attempt to retrofit T-Ranketology to fit last year's results. Next year's algorithm will be adjusted accordingly.

***Update on  2/17/16:

I've revamped the T-Ranketology formula a little bit, as I think the original version was more wishful thinking than predictive. Here are the changes:

1) I removed the current T-Rank as a factor. I put it in originally on the theory that this would be similar to the "Easy Bubble Solver" method which uses Kenpom ranking and RPI ranking to project likely tournament teams. But T-Rank is already a factor in that it is used to project the Elo and RPI ratings that make up the other two factors. Obviously, no one is looking at T-Rank (or Kenpom ranking, really) to determine who gets into the tournament.

2) In place of T-Rank, I've added a "good wins / bad losses" analysis. Teams get 10 points for top 50 wins, 3 points for other top 100 wins and lose 3 points for sub-100 losses, and 6 points for other sub-200 losses. (I may fiddle with these values some -- let me know if you have an idea what they should be instead.) Ideally I would use projected RPI as the source of the rankings, but I'm using current T-Rank because it's much easier for me. [Edit: after further review, it's not hard to use projected RPI, so I've fixed this.] I did this because "good wins" is clearly a big factor in how the people who actually decide this stuff go about deciding it.

The result is that I get a projected bracket that looks a lot more like the consensus. Valpo is no longer projected a 7-seed. Gonzaga no longer seems safely in. Princeton and Yale no longer look like serious bubble teams. Providence is no longer in serious danger of dropping out (although they're still much lower than in other brackets because of T-Rank's expectation that they'll lose some games coming home).

In my opinion, the former T-Ranketology produced a better bracket, but it wasn't as realistic.

*******


As promised, I've been playing around quite a bit with various T-Rank projects this year. Lately, I've added a couple of fun new rankings, including the Q-Rank (based on performance in Tournament Quality Tests), the H-Rank (ratings in last 10 games only) and the E-Rank, a slightly modified "Elo" rating system. I've also put together a program that uses T-Rank to forecast what teams' RPI will be at the end of the regular season.

My marketing people think I'm diluting the T-Rank brand with all these other rankings, but I have to give the people what they want, even if they don't know that it's what they want, yet.

Speaking of which, the latest: T-Ranketology. This is my NCAA bracket projections system. The inputs, in equal measure:

1) Current T-Rank
2) Projected RPI Rank at the end of the regular season
3) Projected Elo Rank at the end of the regular season

Both RPI and Elo are "resume" ranks, and T-Rank is a measure of team strength. The key to understand here is that the RPI and Elo inputs are projections -- that is, I run simulations of the rest of the season to get an average final RPI and Elo. (The T-Rank, however, is treated as a static determinant of team quality.)

So T-Ranketology is not a prediction based on "if the season ended today." It's a projection based on how the season will likely play out, if the current T-Rank is a correct measure of team strength.

One thing to make clear: my projection of the rest of the season does not include conference tournaments. I ain't got time for that.

Finally, I use current T-Rank to pick the autobids. That is, the highest team in the T-Rank is gifted the autobid, on the assumption (not always correct, given seeding effects) that the best team is most likely to win the conference tournament.

As always, this is all in good fun. If you see anything goofy, let me know.

No comments:

Post a Comment